Gaussian processes have become a corner stone for Bayesian inference in many modern applications, especially for inference in high-dimensional spaces with limited amount of training data. The quality of the gaussian process predictive distribution is sensitive to the choice of hyperparameters in its kernel. The predominant workhorse for optimization of the hyperparameters is to use the (derivatives of) the Log Marginal Likelihood (LML). We illustrate a number of situations in which this choice for hyperparameter optimisation is problematic and lead to sub-optimal performance of the resulting gaussian process. We describe a number of different ways to mitigate the sub-optimality, and show that they lead to significantly more robust results for gaussian process accelerated active learning in catalysis-relevant atomic-scale structure simulations. The benefit primarily originate from the improved uncertainty predictions by the gaussian process.

Improving robustness of gaussian processes for active learning of potential energy surfaces

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.